mutual-knn-captures-coarse-categorical-not-fine-grained-alignment
IN premise — summaries/2026/08/24/koepke-2026-back-into-cave-sR-references-chunk-1.md
Created 2026-08-24T17:11:00+00:00
Mutual kNN alignment at k = n/100 remains stable across gallery sizes (WIT-1K through LAION-15M), indicating shared coarse semantic-category structure from overlapping web data, while alignment at fixed small k (1, 10) drops dramatically, indicating absence of fine-grained representational convergence between independently trained unimodal encoders.
Summary
Two independently trained AI models share a rough, category-level map of meaning — they both recognize broad groups like "vehicle" versus "person" — but their internal representations do not actually line up when you look at fine-grained details like specific subtypes or subtle attributes. The shared coarse structure simply reflects that both models saw overlapping web content during training, not any deep representational agreement, so their internal "vocabulary" is comparable only at a very broad level and cannot be assumed to match at the detail level.